JP2008249610A - Method for estimating dynamic characteristic of suspension apparatus for railroad vehicle - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method for estimating a dynamic characteristic of a suspension apparatus for a railroad vehicle component which automatically generates a high-accuracy response estimation model from input/output data of the existing vehicle component measured by a running tester, overcomes the inaccuracy due to an approximation, and efficiently obtains the accurate model. <P>SOLUTION: The method for estimating the dynamic characteristic of the suspension apparatus for the railroad vehicle component inputs relative displacement/speed information at an attachment point of the suspension apparatus of the wheeled railroad vehicle comprising a damper and a spring, outputs a force generated from the suspension apparatus, has a simple dynamics model 13 such as a piecewise linear model and neural network models 12, 14, identifies the nonlinearity of the force generated from the suspension apparatus by use of the neural network models 12, 14 based on the displacement/speed information 11 as teacher data and the force generated from the suspension apparatus, and automatically makes an accurate estimate 15 of a damping force of the suspension apparatus to the inputted arbitrary displacement/speed information 11. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a method for estimating dynamic characteristics of railway vehicle parts, and in particular, a railway vehicle suspension system.

  Conventionally, there have been the following prior arts in this field.

  A characteristic identification example of an air spring and a control system based thereon are proposed using a recurrent or hierarchical neural network model that does not use a dynamic model (see Non-Patent Document 1 below).

  Non-Patent Document 2 below proposes an identification method for separating linear characteristics and nonlinear characteristics by a neural network combining linear neurons and nonlinear neurons.

  FIG. 12 shows such a conventional dynamic process modeling apparatus.

As shown in this figure, a neuron network (101) is connected in series to a transfer element (102) for modeling a nonlinear dynamic process such as a valve or a damper in flow regulation. This neuron network (101) is configured to simulate a static characteristic curve of a process (Patent Document 1 below).
Japanese National Patent Publication No. 11-506553 Identification and Control of Nonlinear Characteristics of Pneumatic Active Suspension for Railway Vehicles by Neural Network, Transactions of the Japan Society of Mechanical Engineers (C), Vol 61, No. 586, pp. 119-124, 1995-6 Experimental identification of nonlinear vibration system by neural network, Transactions of the Japan Society of Mechanical Engineers (C), Vol 67, No. 663, pp. 28-34, 2001-11

  However, in the development of railway vehicles, confirmation by running tests involves great costs and labor, and there are strong restrictions on the use of tracks, etc., so the vehicle motion state in the running state is evaluated by prior simulation, and various vehicle parameters are We are trying to optimize. The dynamic characteristics of mechanical parts such as dampers and air springs are important elements in the simulation. However, because they have strong nonlinearity, conventional linearized and approximated models do not always have sufficient accuracy if the assumptions differ. The behavior could not be reproduced. In addition, trial and error with great effort was necessary to obtain a high-accuracy model.

  The ARX model that performs linear estimation is known as a method for estimating the dynamic characteristics of the test object in a lump. However, the applicable range is limited for components with strong nonlinearity such as dampers, and a wide range of applications can be applied. It was difficult to reproduce the shaking state.

  In view of the above situation, the present invention automatically generates a high-accuracy response estimation model from input / output data of existing vehicle parts measured by a running test apparatus, and eliminates the inaccuracy associated with the above approximation. Another object of the present invention is to provide a method for estimating the dynamic characteristics of a railroad vehicle suspension system that can efficiently obtain an accurate model.

In order to achieve the above object, the present invention provides
[1] In a method for estimating the dynamic characteristics of a railway vehicle component using as input the relative displacement / speed information at the attachment point of a suspension system consisting of a damper and a spring of a railway vehicle with wheels and outputting the generated force of the suspension system, It has a simple dynamic model such as a linear model and a neural network model, and generates the suspension device by the neural network model based on the displacement / velocity information as teacher data and the generated force of the suspension device. It is characterized in that the nonlinearity of the force is identified, and the damping force of the suspension device is automatically estimated accurately for any input of the displacement / velocity information.

  [2] In the method for estimating dynamic characteristics of a railway vehicle suspension system according to [1], the neural network model is a feedforward type.

  [3] In the dynamic characteristic estimation method for a railway vehicle suspension system according to [1] above, prior to learning of the neural network model, the parameters of the dynamic model are automatically optimized from the input / output relationship of the teacher data. The learning load of the neural network model is reduced.

  [4] In the method for estimating dynamic characteristics of a railway vehicle suspension system according to [1], the neural network model is configured by combining linear neurons and nonlinear neurons in an input layer.

  [5] In the dynamic characteristic estimation method for a railcar suspension system according to [1] above, the frequency component ratio (high coherence frequency ratio: HCR) that the coherence between the input and output is equal to or higher than a set value is identified. It is characterized by evaluating the performance of a neural network model.

  [6] In the dynamic characteristic estimation method for a railcar suspension system according to [5] above, over-learning is prevented by performing an abort determination of learning using the evaluation method for input / output data including noise. It is characterized by.

  [7] In the method for estimating the dynamic characteristics of a railcar suspension system according to [1] above, from the experimental data obtained by the multi-degree-of-freedom test apparatus combining a 6-degree-of-freedom motion base and a 6-component force sensor, It is characterized by collectively identifying the responses.

  [8] In the method for estimating dynamic characteristics of a railway vehicle suspension system according to [7] above, as the teacher data of the neural network model, data obtained by extending a range based on displacement / speed information measured by the traveling test apparatus is used. A result obtained by exciting the multi-degree-of-freedom test apparatus is provided.

  [9] In the method for estimating the dynamic characteristics of the railcar suspension system according to [7] above, for generating the teacher data, by giving the driving pattern of the traveling test apparatus the type of the test part and specifications, A drive target signal for the travel test apparatus is automatically generated.

According to the present invention, the following effects can be achieved.
(1) By obtaining a high-accuracy model of a railway vehicle suspension system, simulation accuracy is improved.
(2) Test / identification procedures are automated, and simulation model generation is made more efficient.

  The method for estimating dynamic characteristics of a suspension system for a railway vehicle according to the present invention inputs relative displacement / speed information at the attachment point of the suspension system composed of a damper and a spring of a railway vehicle with wheels, and outputs the generated force of the suspension system. In a method for estimating the dynamic characteristics of railway vehicle parts, a simple mechanical model such as a piecewise linear model and a neural network model are used. Based on the displacement / speed information as teacher data and the generated force of the suspension device Thus, the nonlinearity of the generated force of the suspension device is identified by the neural network model, and the accurate estimation of the damping force of the suspension device is automatically performed for any input of the displacement / speed information.

  Hereinafter, embodiments of the present invention will be described in detail by taking the identification of a railway vehicle damper as an example.

  In order to identify the railway vehicle damper and verify the accuracy, test data of “teacher data” and “verification data” is required. The neural network model of the present invention (see claims 1 to 4) learns the input / output relationship of teacher data, so that the output predicted from the input of the teacher data best matches the output data of the teacher data. Automatically adjust the parameters. This process is called neural network model learning. Depending on the degree of coincidence between the output estimated by applying the measured input / output data different from the teacher data to the network that has completed learning of this neural network model (hereinafter referred to as the identification model), Evaluate the accuracy of the model. This comparison data is called verification data. The verification data is preferably a signal sequence that has no correlation with the teacher data.

  FIG. 1 is a configuration diagram of a 6-degree-of-freedom damper test apparatus for measuring teacher data and verification data used for identification according to the present invention, FIG. 1 (a) is a plan view thereof, and FIG. 1 (b) is a side view thereof. It is.

  In the system of the present invention, teacher data and verification data used for identification are measured by a 6-degree-of-freedom damper test apparatus 1 as shown in FIG. 1 (see claim 7).

  This 6-degree-of-freedom damper test apparatus 1 includes a so-called Stewart-type vibration mechanism 3 that vibrates at the degrees of freedom of up / down, left / right, longitudinal displacement and roll, pitch, and yaw by controlling six actuators 2, It consists of a 6-component force sensor 5 that receives the force generated by the damper 4, and can correspond to all motion modes that appear in the railway vehicle. The command displacement to the Stewart type vibration mechanism 3 is given as a relative displacement and a relative attitude angle between the damper mounting points. 6 is a 6-DOF motion base.

  The excitation pattern for the teacher signal given to the six-degree-of-freedom damper test apparatus 1 is created by expanding the range so as to include the relative displacement and the relative speed actually measured in a running test or the like (see claim 8).

  This situation is shown in FIGS.

  FIG. 2 is a waveform comparison diagram between teacher data and verification data, where the horizontal axis indicates time (s) and the vertical axis indicates displacement (mm). FIG. 3 is a comparison diagram of change ranges of the teacher data and the verification data. The horizontal axis indicates displacement (mm), and the vertical axis indicates speed (mm / s).

  In general, the vibration pattern for verification data uses displacement data measured in a running test.

  These vibration pattern data are automatically generated by giving a damper type and representative specifications to the drive pattern generation program (see claim 9).

  An example of creating the teacher data is shown in FIG.

  In FIG. 4, A is a damper selection information input unit for selecting a damper, B is a damper characteristic selection information input unit for selecting a damper characteristic, C is a past data reading information input unit for reading past data, D Is a waveform comparison display unit indicating the Y-axis displacement between the teacher data and the verification data, and E is a comparison display unit for the change range of the teacher data and the verification data.

  The test damper 4 is vibrated with these vibration patterns, the generated damping force is measured by the 6-component force sensor 5, and after AD conversion, the 6-DOF motion calculated from the displacement of the 6 actuators 2 Along with the displacement data of the reference point of the base 6, it is recorded in a file as teacher data or verification data.

  The automatic identification program uses the parameters of the neural network model so that the evaluation function for evaluating the error of the generated force estimated from the displacement data is below the allowable value based on the data file of the teacher data and the given damper specifications. Adjust automatically.

  An example of this network structure is shown in FIG.

  FIG. 5 is a block diagram of a neural network model according to the present invention.

  In this figure, 11 is displacement / velocity information, 12 is a nonlinear part composed of nonlinear neurons □ receiving displacement / velocity information 11, 13 is a dynamics model (mechanical model) in which displacement information is taken in, 14 is from the nonlinear part 12 A linear unit composed of linear neurons ○ receives an output from the mechanical model 13, 15 is an output unit of the estimated damping force from the linear unit 14, and 16 is an estimated attenuation of the estimated damping force from the output unit 15. The comparison unit 18 compares the force output with the learning data 17 of the actually measured damping force, 18 is a learning rule obtained from the comparison unit 16, and the learning rule 18 is a neural network model (nonlinear unit 12 and linear unit 14). Will be adjusted.

  Examples of this network include (1) a simple mechanical model 13 to be identified and (2) a neural network model (nonlinear part 12 and linear part 14), and (3) arbitrary displacement / velocity information 11 A damping force is estimated for (input) (see claim 1).

  This network is feedforward (see claim 2) because the output is not returned to the input. Further, the transfer function of the neuron used therein has a structure in which linear neurons and nonlinear neurons are used together (see claim 4). In this example of the neural network model, the following tangent sigmoid function formula (1) is used as the nonlinear neuron, but other functions may be used.

A recurrent type in which an output is returned to an input is known for a feedforward type network configuration. The recurrent type is said to be suitable for estimating input / output relationships including past histories. There is a risk of divergence if it is significant.

  In this respect, the feedforward type is considered to be advantageous in terms of stability because the output is not recirculated even if it receives such a large input, because the output does not return. However, in this example, in order to make the neural network model learn dynamics, velocity information is created from the displacement information, and input data is given as a set of displacement and velocity.

  As an evaluation function for performing learning, for example, the following average error sum-of-squares expression (2) is suitable. As a learning rule, Newton's method or the like can be used.

In this example, a spring damper series model as shown in FIG. 6 was used as a mechanical model (see claim 1).

  FIG. 6 shows a spring damper series model, k is a spring, and c is a damper.

  Further, a piecewise linear model as shown in FIG. 7 is used in which the damping coefficient is switched from a piston speed in consideration of the actual damper characteristics.

  FIG. 7 shows a piecewise linear model, in which the horizontal axis represents the piston speed and the vertical axis represents the generated force. In this dynamic model, it is necessary to designate four parameters: (1) damping coefficient 1, (2) damping coefficient 2, (3) characteristic change point, and (4) spring coefficient. These parameters can be determined by performing linear optimization based on the input / output relationship of the teacher damper as preprocessing (see claim 3). An example of this is shown in FIG.

  FIG. 8 is an explanatory diagram of the optimization of the pre-mechanical model.

  In this figure, 21 is the displacement information, 22 is the dynamic model, 23 is the identification target, 24 is the estimated spring strain amount from the mechanical model 22 and the spring strain amount from the identification target 23, and the parameters are optimized. An adder that outputs a value and feeds back to the mechanical model 22.

  Therefore, optimization of the pre-mechanical model is performed by, for example, the above-described four parameters, that is, (1) damping coefficient 1, (2) damping coefficient 2, (3) characteristic change point, and (4) spring coefficient. On the other hand, it can be executed by applying the Newton method using the average error square sum of the measured damping force and the estimated damping force as an evaluation function.

In this example, an excellent estimation result was obtained from the neural network by setting the output of the mechanical model 22 to the amount of spring strain xx ₁ .

  Since the teacher data and the verification data include noise, the goodness of the identification model cannot always be measured only by the sum of squared errors of the measured damping force and the estimated damping force. For example, in the situation where the noise component is included in the teacher data, if an extremely small mean error sum of squares is specified as a learning target, the neural network model is adjusted to express even fluctuation due to noise, and is not necessarily identified. It will not show the characteristics of the target. Such a state is called “overlearning”.

  The over-learning identification model is significantly less adaptable to different patterns of input, and the learned data shows good estimation, but the other input data shows estimation results that are significantly different from actual measurements. For this reason, there is a need for an index for evaluating that the identification model reflects the nature of the teacher object. As this index, “high coherence frequency ratio” was proposed (see claim 5).

  This underlying coherence is calculated by the following formula (3) and is an index indicating the strength of the relationship between two signals for each frequency component. Generally, a strong correlation is obtained in a range having a value of 0.8 or more. There is. Since causality is expected between the input and output to be identified, it is expected that there will be a certain frequency band with high coherence between the input and output.

FIG. 9 shows the definition of the high coherence frequency ratio.

  The ratio of the total output power to the total output power in the frequency band where the coherence between input and output (displacement / velocity, damping force) exceeds the threshold is defined as “high coherence frequency ratio (abbreviated as HCR)”. . This amount indicates the strength of the relationship between the input and output of the identification result. In this example, 0.8 was used as the threshold value.

  The influence of noise acts to lower the correlation between input and output, so when overlearning occurs, the correlation between input and output decreases. For this reason, it can be determined whether it is in an overlearning state by evaluating the transition of HCR (refer to Claim 6).

  FIG. 10 is a diagram showing the relationship between the mean square error and the high coherence frequency ratio (HCR).

In this figure, when noise is not included in the output, the HCR increases as the mean square error decreases and approaches 100%. However, when noise is included, it decreases when exceeding the appropriate value. . This indicates that the network is adjusted so as to be tuned to a noise component unrelated to the response to be identified, and it can be said that it is in an overlearning state. In this example, the HCR was maximized when the mean square error was around 5 × 10 ⁻⁴ to 10 ⁻³ .

  FIG. 11 shows the result of comparing the accuracy of the identification model generated as described above with the actual vibration pattern by the running test. Compared with the ARX model, which is a typical linear identification method, the average error square sum is reduced by 33% in the case of the left and right motion damper (FIG. 11A), and in the case of the yaw damper [FIG. 11B]. The usefulness of this technique was demonstrated by a 76% reduction.

  As described above, we developed a system that automatically identifies the characteristics of nonlinear dampers and air springs, and confirmed that high-precision identification is possible using a neural network model.

  In addition, we were able to obtain a method for improving the simulation accuracy in conjunction with the test equipment.

  In addition, this invention is not limited to the said Example, Based on the meaning of this invention, a various deformation | transformation is possible and these are not excluded from the scope of the present invention.

  The automatic characteristic estimation method for railway vehicle parts of the present invention can be used as a dynamic characteristic estimation method for mechanical parts such as a railway vehicle damper device.

It is a block diagram of a 6-degree-of-freedom damper test apparatus for measuring teacher data and verification data used for identification according to the present invention. It is a waveform comparison diagram of teacher data and verification data. It is a comparison figure of the change range of teacher data and verification data. It is a figure which shows the example of preparation of teacher data. It is a block diagram of the neural network model concerning this invention. It is a figure which shows a spring damper serial model. It is a figure which shows a piecewise linear model. It is explanatory drawing of optimization of a pre-mechanical model. It is explanatory drawing of a high coherence frequency ratio. It is a figure which shows the relationship between a mean square error and a high coherence frequency ratio (HCR). It is a figure which shows the identification result of a left-right dynamic damper and a yaw damper. It is a figure which shows the modeling device of the conventional dynamic process.

Explanation of symbols

1 6-degree-of-freedom damper test device 2 6 actuators 3 Stewart-type vibration mechanism 4 Damper 5 test 6-force sensor 6 6-degree-of-freedom motion base A Damper selection information input section for selecting damper B Select damper characteristics Damper characteristic selection information input section to be performed C Past data read information input section to read past data D Waveform comparison display section showing displacement of Y axis of teacher data and verification data E Comparison display of change range of teacher data and verification data Part 11 Displacement / velocity information 12 Non-linear part composed of non-linear neurons □ 13 Dynamics model (mechanical model) that incorporates displacement information
14 Linear part composed of linear neurons ○ 15 Estimated damping force output part 16 Comparison part 17 Measured damping force learning data 18 Learning rule 21 Displacement information 22 Dynamic model 23 Identification target 24 Adder part

Claims (9)

In the method for estimating the dynamic characteristics of railway vehicle parts, the relative displacement and speed information at the attachment point of the suspension device composed of a damper and a spring of the railway vehicle with wheels is input, and the generated force of the suspension device is output.
(A) a simple mechanical model such as a piecewise linear model;
(B) a neural network model;
(C) Based on the displacement / velocity information serving as teacher data and the generated force of the suspension device, non-linearity identification of the generated force of the suspension device is performed by the neural network model. A method for estimating a dynamic characteristic of a suspension device for a railway vehicle, which automatically estimates an accurate damping force of the suspension device with respect to an input.   The dynamic characteristic estimation method for a railway vehicle suspension system according to claim 1, wherein the neural network model is a feedforward type.   The dynamic characteristic estimation method for a railway vehicle suspension system according to claim 1, wherein, prior to learning of the neural network model, parameters of the mechanical model are automatically optimized from an input / output relationship of the teacher data, and the neural network A method for estimating a dynamic characteristic of a suspension system for a railway vehicle, wherein the learning load of the model is reduced.   2. The dynamic characteristic estimation method for a railway vehicle suspension system according to claim 1, wherein the neural network model is configured by combining linear neurons and nonlinear neurons in an input layer. Method.   The dynamic characteristic estimation method for a railway vehicle suspension system according to claim 1, wherein a frequency component ratio (high coherence frequency ratio: HCR) at which the coherence between the input and output is equal to or greater than a set value is A method for estimating a dynamic characteristic of a suspension for a railway vehicle, characterized by evaluating performance.   6. The dynamic characteristic estimation method for a railway vehicle suspension system according to claim 5, wherein over-learning is prevented by performing an abort determination of learning using the evaluation method for input / output data including noise. A method for estimating the dynamic characteristics of a suspension system for a railway vehicle.   The dynamic characteristic estimation method for a railway vehicle suspension system according to claim 1, wherein responses of six degrees of freedom are collectively obtained from experimental data obtained by a multi-degree-of-freedom test apparatus combining a six-degree-of-freedom motion base and a six-component force sensor. The dynamic characteristic estimation method of the suspension system for rail vehicles characterized by identifying as follows.   8. The method for estimating dynamic characteristics of a railway vehicle suspension system according to claim 7, wherein the multi-freedom data is obtained by expanding the range based on displacement / speed information measured by the travel test apparatus as teacher data of the neural network model. A method for estimating the dynamic characteristics of a suspension device for a railway vehicle, characterized in that a result obtained by applying vibration with a degree test device is provided.   8. The method for estimating dynamic characteristics of a railway vehicle suspension system according to claim 7, wherein the driving pattern of the traveling test apparatus is automatically given by providing a type and specification of a test part for generating the teacher data. A method for estimating a dynamic characteristic of a suspension device for a railway vehicle, comprising generating a drive target signal for a traveling test apparatus.